【作者】 郭峰；

【导师】 杨静；

【作者基本信息】 太原理工大学 ， 电子与通信工程， 2016， 硕士

【摘要】 图像融合是对包含同一场景信息的多幅源图像进行处理并将各源图像中的有用信息进行提取,最终利用这些信息构成具有完整场景信息的融合图像,使融合后的图像具有更好的机器视觉效果和人类视觉效果,以便后续的研究和应用能够顺利进行。目前,图像融合主要在空间域和变换域两个方面被研究,基于空间域的融合方法是利用特定方法直接对图像像素进行计算融合;基于变换域的融合方法首先要对源图像进行变换处理,以得到能够体现图像特征的变换系数,然后通过特定的融合规则将各源图像的变换系数进行融合,最后将其进行逆变换便可得到最终的融合图像。如何有效并且精确的对源图像的有用信息进行检测是图像融合的关键,对多种图像信息提取方法进行研究后,本文提出了一种基于可控滤波和超像素分割相结合的图像融合方法。可控滤波器对图像的边缘细节信息具有较强的敏感性,因此本文利用可控滤波器对图像的边缘细节等高频信息进行检测提取,而对于图像的结构信息,即低频成分,本文利用超像素分割的方法进行提取,最后利用本文所提出的融合规则将这些信息进行有效融合,从而使最终的融合图像包含了来自于不同源图像的有用信息。本文的主要工作与创新:(1)对有关图像融合的知识进行了概括性的介绍;(2)阐述了可控滤波器的设计思想,并对其设计原理进行了介绍,同时利用实验比较的方法验证了可控滤波器在图像边缘细节信息提取方面的有效性。(3)对基于熵率的超像素分割方法进行研究,总结了利用熵率的图上随机游走模型对图像进行分割的原理,并且介绍了利用贪心算法解决超像素分割问题的具体过程。(4)在将可控滤波器以及超像素分割应用到图像融合的过程中,提出了方向响应差的概念以及基于选点与选面的两种融合规则,使得本文算法更加合理,同时提高了算法的稳定性。(5)设计制作了便于实验研究的GUI(图形用户界面),使得方法选择、参数分析等实验环节更加便捷。同时,利用多组实验图像从主观评价和客观评价两个方面对本文算法的融合结果与其它经典算法的融合结果进行比较分析,验证了本文算法的合理性以及有效性;在抗噪性分析中,通过对源图像加入泊松噪声、高斯噪声、椒盐噪声以及乘性噪声的实验验证了本文算法的稳定性;最后,利用实验对算法中的可变量进行了参数分析,并总结了它们对算法融合结果的影响规律。更多还原

【Abstract】 Image fusion process source images which contain the same scene information and extract useful information form these source images, the final fused image is constituted with these useful information so that it has better human visual effect and machine visual effect. At present, image fusion method is divided into two categories: the fusion method based on the spatial domain and the fusion method based on the transform domain. Images are fused directly on the gray scale image pixel space, if the fusion method based on the spatial domain is used; the fusion method based on the transform domain: Firstly, each source image is transformed separately. Secondly, transformed coefficients are fused according to a specific rule. Finally, the final fusion image is obtained by inverse transforming of transformed coefficients.The key of image fusion is that how to detect the useful information of source images effectively and accurately. After several methods about image information extraction were researched, the image fusion based on steerable filters and super-pixel segmentation was proposed and it has been studied. The edge and high-frequency details information of source images were detected and extracted by steerable filters, because the steerable filters had a strong sensitivity for edge information of image,and structure and low-frequency componentsinformation were detected and extracted by super-pixel segmentation. Finally,these information were fused effectively by the proposed fusion rules so that the final fusion image contained useful information which from different source images.The main work in this dissertation is listed as follows:(1) Knowledge about image fusion has been a general introduction;(2) The design idea of the steerable filter was described, and its design principle was introduced. At the same time, the effectiveness of the steerable filter for image edge detail information extraction was verified by experiments.(3) The super-pixel segmentation method based on entropy rate was researched, summarized the image segmentation principle with the random walks model on graphs of entropy rate, and introduced to the specific process of super-pixel segmentation problem were solved by greedy algorithm.(4) Steerable filter and super pixel segmentation were applied to the process of image fusion, and the direction response difference and two fusion rules which based on selecting point and selecting surface were proposed, so that the proposed algorithm is more reasonable, and improving the stability of the proposed algorithm.(5) Graphical User Interface was designed which made more convenient in experiments, such as selection of fusion methods and analysis of parameters.At the same time, the proposed algorithm was reasonable and effective that was verified by several experiments which contained two aspects: subjectiveevaluation based on visual and objective evaluation based on evaluation indexes,and the fusion result of proposed algorithm and the fusion results of other classical algorithms were compared and analyzed. Stability of the proposed algorithm was verified by source images which were fused with Poisson noise and Gaussian noise, salt and pepper noise and multiplicative noise. Finally,variable parameters of the proposed algorithm were analyzed by experiments,and the law, the influence of parameters on the result of the proposed algorithm,was summarized.更多还原